Running a memory profiler on simulations using neurodamus to create the input files for coreneuron in multiple cycles showed that memory allocated in:

accumulated across cycles. The function is called from write_nrnthread, with the following context:

        double* data = NULL;
        nrnthread_dat2_mech(..., data, ...);

        if (nrn_is_artificial_[type]) {
            delete[] data;
        }

looking into nrnthread_dat2_mech we find the following structure:

int nrnthread_dat2_mech(..., double*& data, ...) {
    bool copy = data ? true : false;

    // As the NEURON data is now transposed then for now always create a new
    // copy in the format expected by CoreNEURON -- TODO make sure this is
    // freed, then later TODO remove the need for this entirely
    if (!copy) {
        data = new double[n * sz];
    }
    // Don't give away/store `data` anywhere.

In summary when called from write_nrnthread:

1. The argument data passed to nrnthread_dat2_mech is unconditionally NULL.
2. Memory will be allocate to data unconditionally.
3. Memory is only released when nrn_is_artificial_[type] which seems to only sometimes hold.

The immediate fix is to simply release the memory since it doesn't seem to be used afterwards. Profiling the same run shows the expected reduced memory consumption.